KLI

FAULT DIAGNOSIS OF ROTATING MACHINES BASED ON ADVANCED SIGNAL PROCESSING METHODS AND DATA-DRIVEN ARTIFICIAL INTELLIGENCE TECHNIQUES

Metadata Downloads
Abstract
Robust condition assessment and predictive maintenance of industrial equipment is a valuable and ever-growing research field. The core of this field is relying on accurate condition monitoring, fault detection, and fault identification in industrial equipment. Fault detection and identification can be applied to a huge variety of industrial components such as rotating machines, blade turbines, industrial robot manipulators, motors, pipelines, gearboxes, spherical tanks, pressure vessels, etc. This work focuses on problems of diagnosing blade rub-impact faults of various intensities in turbines. Rubbing faults, also called in literature as rub-impact faults, represent a class of highly nonlinear and nonstationary mechanical faults that may appear in different rotating machines in general, and turbomachinery, specifically. These faults are caused by the interaction between rotating and stationary parts of rotating machines. The rubbing phenomenon may appear in the industrial system not only as a single isolated fault but as the consequence of the presence of other mechanical faults (mostly, rotor-related faults), which creates challenges for fault diagnosis and analysis. If the presence of the rubbing process in the rotor system is not detected and its severity is not identified in time, it can lead to the failure of the whole mechanical system (i.e., turbine). The failure of the system results in unexpected downtimes and great economic losses. Moreover, any mechanical fault appearing in industrial equipment can become a threat to the safety of the people working in the facility. The condition monitoring of turbines and the diagnosis of blade rub-impact faults can be achieved using different types of techniques. Therefore, in this dissertation, the approaches based on analyzing vibration signals using advanced signal processing and data-driven artificial intelligence algorithms are developed that can accurately identify and diagnose blade rub-impact faults of various severity levels.
Due to the cost-effectiveness and reliability requirements, fault diagnosis of the industrial equipment using vibration acceleration signals has been a key direction of research over the past several decades that gave us a variety of methods. However, due to the complex nonstationary and nonlinear nature of rubbing faults, the application of the conventional time- and frequency-domain signal analysis techniques for examining and extracting discriminative fault features of rubbing processes appeared to be not effective. Furthermore, another condition that should be considered when diagnosing rub-impact faults is the fact that the rubbing phenomenon can be induced by other mechanical faults appearing in the system, such as bearing failures, shaft imbalance, etc. In these cases, it is favorable to apply the time-frequency analysis (TFA) methods for vibration signal decomposition. Moreover, it is crucial to improve the quality of the rubbing signal decomposition as well as determine, which frequency components of the signal (i.e., sub-bands) are exactly related to the mechanical fault being investigated. For this, the hybrid feature extraction based on the advanced TFA decomposition technique called ensemble empirical mode decomposition (EEMD) in conjunction with the autonomous fusion of optimal intrinsic mode functions (IMFs) containing harmonics of blade rub-impact is proposed in this dissertation. This dissertation also introduces the ratio between degree-of-frequency presence (DFP) of blade rub-impact fault-related harmonics and Kullback-Leibler divergence (KLD) as a specific criterion for autonomous selection of subset containing optimal IMFs related to rubbing faults. The experimental results indicated that the signals reconstructed using selected optimal IMFs related to the rubbing process contained less high-frequency noise and excluded the frequency harmonics not related to rub-impact faults. The hybrid feature pool extracted from the reconstructed signals and used as the input to one-against-all multiclass support vector machines (OAAMCSVM) classifier demonstrated high discriminative properties not only for detecting but also for diagnosing coupling rotor imbalance and blade rub-impact faults of various intensity levels.
Intrinsic mode fusion appeared to be an important procedure that should be used for extracting valuable information about the mechanical faults being investigated when the iterative TFA decomposition techniques, such as empirical mode decomposition (EMD) and EEMD, are used for analyzing complex nonstationary signals. Despite DFP/KLD-based optimal IMF selection algorithm demonstrated its high capabilities of selecting intrinsic modes that contain valuable information about rubbing processes for vibration signal reconstruction, the application of this method may lead to the selection of non-informative intrinsic components along with informative ones in some extremal cases when computing the objective function value. Therefore, in this thesis, the originally proposed DFP/KLD-based optimal intrinsic mode fusion algorithm has been improved with the adaptive thresholding and objective function normalization techniques to increase the robustness of this selection method and reduce the possibility of selecting non-informative IMF components. The experimental results demonstrated that the vibration signals containing coupling rotor imbalance and blade rubbing faults reconstructed using the set of optimal IMFs selected by the improved selection algorithm include less high-frequency noise in comparison with the originally proposed selection framework while still preserving the valuable features (frequency harmonics) evidencing about rubbing process ongoing in the system. Furthermore, the results show the improvements in terms of fault classification accuracy in comparison with the original technique.
Machine learning (ML)–based condition monitoring and fault diagnosis of a rotary machine, such as turbines, are inevitable given the complexity of many problems. The operation of turbines is dependent on the health condition of its blades that operate in a hostile and high-stress environment. Moreover, due to harsh working conditions and precise requirements for clearance between stator and rotor blades, blade rubbing fault becomes a highly nonlinear type of mechanical faults in which several physical phenomena, such as vibration, friction, thermal effects, and stiffness are involved simultaneously. These factors affect and deteriorate the performance of conventional fault diagnosis models and make the fault diagnosis process more challenging that frequently leads to the incapability of the signal-based fault identification frameworks to maintain their high performance when the operating conditions of rotor systems, such as rotating speed and load, change. In this regard, the system-based data-driven technique from the family of artificial intelligence (AI) methods is presented for detection and identification of coupling rotor imbalance and blade rubbing faults. The data-driven system-based fault identification framework is developed based on approximation of the nonlinear function of the rotor system by deep undercomplete denoising autoencoder (DUDAE) and fault classification by a deep neural network (DNN). The contributions of this work are as follows. First, the proposed DUDAE learns the approximation of the nonlinear function of the rotor system by vibration signals collected under normal operating conditions, i.e., when neither rotor imbalance nor blade rub-impact faults present. Next, the residual signal is computed as the difference between the unknown vibration signal coming from the rotor system and its estimate provided by trained DUDAE. Finally, the DNN is applied to perform fault classification based on the generated residual signals. The experimental results demonstrated that the generated residual signals are sensitive to the rotor system degradation allowing for high fault classification accuracy when they are used as features for determining the state of the system. The obtained results of this study are compared with current state-of-the-art deep learning-based fault diagnosis techniques, which demonstrates the advantages of the proposed model.
Multivariate signal analysis and data fusion from multiple sensors are important topics in the field of condition monitoring that can significantly increase the performance of fault diagnosis frameworks. The simultaneous analysis of multivariate signal allows for a precise investigation of different processes ongoing in the rotor systems. With the increase of computational complexity of TFA approaches when dealing with multivariate signals, the deep learning-based techniques that perform representation learning drew the attention of researchers and industrial specialists when addressing the problem of fault diagnosis using multivariate signals. In current work, a mechanism has been developed that can recognize and identify coupling rotor imbalance and blade rub-impact faults of various intensity levels by using multivariate vibration signals, i.e., the vibration signal collected by several channels of multiple sensors. To reduce the computational complexity and accelerate fault identification procedure, a fault diagnosis model is developed based on vibration signal resampling concerning fundamental frequency, envelope power spectra analysis of multivariate rub-impact fault signals, and a tiny multivariate-one-dimensional convolutional neural network (ModCNN). The contributions of this work are as follows. First, the multivariate vibration signals in the time domain are resampled with overlap using the fundamental frequency to ensure that these samples contain valuable information obtained during each revolution of the rotor. Second, the envelope power spectra of the resampled signals are computed to create discriminative patterns for the further representation learning task. Finally, ModCNN based on adaptive moment estimation optimization function (Adam) is proposed to extract local features from the resampled multivariate signals, hence reducing the overhead of feature extraction and selection. Adam optimization is computationally efficient and is well suited to problems with big datasets. The proposed framework has been evaluated by the vibration dataset collected using the testbed provided by UIAI Lab. and the obtained results of this study are compared with the current state-of-the-art machine and deep learning techniques used in the field of fault diagnosis of rotary machines, which demonstrate the superiority of the proposed model and of multivariate signal analysis.
Solutions from the field of control theory are widely used in industry for assessing the health condition of engineering systems because, in general, they are more robust and reliable in comparison with data-based (signal-based) techniques, since the operation of those algorithms is based on the system dynamics. For fault detection and diagnosis, first, these solutions require designing the observers which are used for improving the accuracy of system identification. One of the most frequently used observers in the industry is the linear-based observation technique. Linear-based observers have the advantage of the design and implementation simplicity along with flexibility; however, they suffer the challenge of lack of robustness and reliability, especially, when they are applied to resolve the problems of nonlinear nature. The problems of linear observers can be generally addressed in two ways: 1) design the nonlinear-based observation technique or 2) improve the performance of linear observers using other types of algorithms, such as artificial intelligence-based techniques. Apart from the advantages of nonlinear-based observers such as robustness and reliability, designing these types of techniques is a complex problem. Thus, enhancing the performance of linear observer when applied to resolve a fault diagnosis problem of nonlinear nature by applying the artificial intelligence-based solution, and thus, introducing a hybrid fault identification technique, can be a suitable solution to this problem. In the current work, the hybrid approach for diagnosing blade rub-impact faults (a nonlinear mechanical fault) using a deep learning-based observation technique is introduced for addressing the problem mentioned above. A hybrid fault diagnosis model is developed based on a linear observation technique called autoregressive with eXogenous input Laguerre (ARX-Laguerre) proportional-integral observer (ARXLPIO) that is used for rub signal estimation and scalable deep neural network (S-DNN) which is utilized for improving the fault identification performance of ARXLPIO. The specific contributions of this work are as follows. First, the ARXLPIO is used to estimate the blade rub-impact signals using the principles of system identification and system estimation. Next, the S-DNN the architecture of which is autonomously designed utilizing the novel proposed algorithm is used for improving the fault identification performance of ARXLPIO when applied to nonlinear problems. The experimental results show that the proposed hybrid approach for diagnosing coupling shaft imbalance and blade rub-impact faults demonstrated high fault classification accuracy and outperformed the referenced methodologies used for the comparison. Furthermore, the simplicity of designing the proposed framework makes it favorable for usage in industrial applications.
Author(s)
프로스비린 알렉산데르
Issued Date
2021
Awarded Date
2021-02
Type
Dissertation
URI
https://oak.ulsan.ac.kr/handle/2021.oak/5940
http://ulsan.dcollection.net/common/orgView/200000369467
Alternative Author(s)
Alexander E. Prosvirin
Affiliation
울산대학교
Department
일반대학원 전기전자컴퓨터공학과
Advisor
Jong-Myon Kim
Degree
Doctor
Publisher
울산대학교 일반대학원 전기전자컴퓨터공학과
Language
eng
Rights
울산대학교 논문은 저작권에 의해 보호받습니다.
Appears in Collections:
Computer Engineering & Information Technology > 2. Theses (Ph.D)
Authorize & License
  • Authorize공개
Files in This Item:

Items in Repository are protected by copyright, with all rights reserved, unless otherwise indicated.